Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

fix: bound the data URL metadata walk in tab-chat persistence

Codex release-state testf390456ddacf
diff --git a/src/chrome/src/ui/tab-chat-persistence.js b/src/chrome/src/ui/tab-chat-persistence.js
index 281497973..5b8397448 100644
--- a/src/chrome/src/ui/tab-chat-persistence.js
+++ b/src/chrome/src/ui/tab-chat-persistence.js
@@ -40,6 +40,17 @@ function isBase64Code(code) {
|| code === 61;
}
+function isMediaParameterCode(code) {
+ return isMimeTypeCode(code)
+ || code === 33
+ || code === 37
+ || code === 39
+ || code === 42
+ || code === 61
+ || code === 95
+ || code === 126;
+}
+
function findImageDataUrl(source, start) {
const firstCode = IMAGE_DATA_URL_PREFIX.charCodeAt(0);
for (let index = start; index <= source.length - IMAGE_DATA_URL_PREFIX.length; index++) {
@@ -56,9 +67,13 @@ function imageDataPayloadEnd(source, marker) {
if (cursor === mimeStart) return -1;
// Walk the short metadata section without applying a regexp to the image
- // bytes. A comma before `;base64,` means this is not a base64 data URL.
- while (cursor < source.length) {
- if (source.charCodeAt(cursor) === 44) return -1;
+ // bytes. Only `;name=value` media-type parameters may sit between the MIME
+ // type and `;base64,`, so the walk stops at the first character that cannot
+ // belong to one (a quote, a tag boundary, whitespace, or the comma that
+ // starts a non-base64 payload). Without that bound the scan would run past
+ // the attribute it started in and swallow every character up to the next
+ // image in the document.
+ while (cursor < source.length && source.charCodeAt(cursor) === 59) {
if (matchesAsciiIgnoreCase(source, cursor, BASE64_DATA_URL_MARKER)) {
cursor += BASE64_DATA_URL_MARKER.length;
const payloadStart = cursor;
@@ -66,6 +81,7 @@ function imageDataPayloadEnd(source, marker) {
return cursor > payloadStart ? cursor : -1;
}
cursor++;
+ while (cursor < source.length && isMediaParameterCode(source.charCodeAt(cursor))) cursor++;
}
return -1;
}
diff --git a/src/firefox/src/ui/tab-chat-persistence.js b/src/firefox/src/ui/tab-chat-persistence.js
index d214620b9..eae3d143f 100644
--- a/src/firefox/src/ui/tab-chat-persistence.js
+++ b/src/firefox/src/ui/tab-chat-persistence.js
@@ -40,6 +40,17 @@ function isBase64Code(code) {
|| code === 61;
}
+function isMediaParameterCode(code) {
+ return isMimeTypeCode(code)
+ || code === 33
+ || code === 37
+ || code === 39
+ || code === 42
+ || code === 61
+ || code === 95
+ || code === 126;
+}
+
function findImageDataUrl(source, start) {
const firstCode = IMAGE_DATA_URL_PREFIX.charCodeAt(0);
for (let index = start; index <= source.length - IMAGE_DATA_URL_PREFIX.length; index++) {
@@ -56,9 +67,13 @@ function imageDataPayloadEnd(source, marker) {
if (cursor === mimeStart) return -1;
// Walk the short metadata section without applying a regexp to the image
- // bytes. A comma before `;base64,` means this is not a base64 data URL.
- while (cursor < source.length) {
- if (source.charCodeAt(cursor) === 44) return -1;
+ // bytes. Only `;name=value` media-type parameters may sit between the MIME
+ // type and `;base64,`, so the walk stops at the first character that cannot
+ // belong to one (a quote, a tag boundary, whitespace, or the comma that
+ // starts a non-base64 payload). Without that bound the scan would run past
+ // the attribute it started in and swallow every character up to the next
+ // image in the document.
+ while (cursor < source.length && source.charCodeAt(cursor) === 59) {
if (matchesAsciiIgnoreCase(source, cursor, BASE64_DATA_URL_MARKER)) {
cursor += BASE64_DATA_URL_MARKER.length;
const payloadStart = cursor;
@@ -66,6 +81,7 @@ function imageDataPayloadEnd(source, marker) {
return cursor > payloadStart ? cursor : -1;
}
cursor++;
+ while (cursor < source.length && isMediaParameterCode(source.charCodeAt(cursor))) cursor++;
}
return -1;
}
diff --git a/test/run.js b/test/run.js
index bb153e84b..8337b6239 100644
--- a/test/run.js
+++ b/test/run.js
@@ -43802,6 +43802,30 @@ test('tab-chat persistence strips large and mixed-case image data URLs without c
assert.equal(chromeResult, firefoxResult, 'chrome/firefox image compaction diverged');
});
+test('tab-chat persistence keeps markup between a bare image MIME mention and a later payload', () => {
+ const input = [
+ '<div>paste as data:image/png here</div>',
+ '<b>kept</b>',
+ `<img src="data:image/gif;base64,${'A'.repeat(64)}">`,
+ ].join('');
+ const expected = [
+ '<div>paste as data:image/png here</div>',
+ '<b>kept</b>',
+ `<img src="${TabChatPersistenceCh.TRANSPARENT_PIXEL_PNG_DATA_URL}">`,
+ ].join('');
+
+ for (const [label, persistence] of [
+ ['chrome', TabChatPersistenceCh],
+ ['firefox', TabChatPersistenceFx],
+ ]) {
+ assert.equal(
+ persistence.stripImagePayloadsForPersist(input),
+ expected,
+ `${label}: markup between a bare MIME mention and a later image payload was swallowed`,
+ );
+ }
+});
+
test('tab-chat persistence never evicts other chats when shared quota remains exhausted', async () => {
for (const [label, persistence] of [
['chrome', TabChatPersistenceCh],